AMD Detection

Answering Machine Detection (AMD) automatically determines whether a call was answered by a human or a voicemail system. This allows the platform to route human-answered calls to agents instantly while automatically leaving voicemail messages on machines — avoiding wasted agent time on voicemail greetings.

Two-Layer Detection

Dialer.CX uses a two-layer approach for maximum accuracy, combining fast heuristic detection with AI-powered deep analysis.

Layer 1: Fast Detection

Rapid initial classification based on audio patterns:

• Analyzes greeting duration, silence gaps, and word patterns
• Results in under 1 second for minimal call connection delay
• High confidence results are acted on immediately
• Uncertain results are escalated to Layer 2

Layer 2: AI Classification

AI-powered deep analysis for ambiguous cases:

• Uses your configured AI provider for advanced audio classification
• Analyzes speech cadence, vocabulary, and greeting structure
• Higher accuracy for edge cases like short human greetings or unusual voicemail messages
• Provider and confidence threshold configurable per campaign

Note

Layer 2 AI classification is optional and can be enabled per campaign. It adds a small amount of latency but significantly improves accuracy for difficult cases.

Configuration

Navigate to Admin > AMD with four management tabs.

Dashboard Tab

Real-time AMD performance analytics:

• Campaign-level AMD accuracy statistics with human vs. machine breakdown
• Daily accuracy trends charted over time
• False positive and false negative rates per campaign
• Performance comparison across campaigns to identify tuning opportunities

Campaigns Tab

Per-campaign AMD tuning parameters:

Setting	Description
Enable AMD	Turn detection on/off per campaign
AI Classification	Enable Layer 2 AI analysis for uncertain cases
Confidence Threshold	Minimum confidence score for machine classification (lower = more aggressive)
Initial Silence	Maximum silence duration before greeting starts (ms)
Greeting Duration	Maximum greeting length before flagging as machine (ms)

Tip

Start with default thresholds and adjust based on the Dashboard accuracy data. Each campaign may need different tuning depending on the demographics and call patterns.

Voicemail Drop Tab

Pre-recorded messages for automatic delivery to answering machines:

• Upload audio files directly or generate via TTS using the built-in text-to-speech engine
• Assign different voicemail messages per campaign for targeted messaging
• Automatic delivery when voicemail is detected — no agent involvement needed
• Track delivery success rates

Fingerprints Tab

Known voicemail greeting pattern library:

• Carrier voicemail pattern database (e.g., “The person you are calling is not available…”)
• Improves Layer 1 detection accuracy by matching known patterns
• Auto-learning from confirmed classifications builds the library over time
• Reduces false positives by recognizing carrier-specific greeting formats

How AMD Integrates with Campaigns

The AMD flow for each outbound call:

1. Call connects — audio stream begins
2. Layer 1 analysis — fast heuristic classification runs immediately
3. High confidence human — call routes to the next available agent
4. High confidence machine — voicemail drop message plays (if configured)
5. Uncertain — Layer 2 AI classification runs (if enabled)
6. Layer 2 result — routes to agent or drops voicemail based on AI classification

Caution

Overly aggressive AMD settings (low confidence thresholds) can cause false positives — real humans classified as machines. Monitor your false positive rate in the Dashboard tab and adjust thresholds accordingly.